Neurosurgery Resident, PGY-5 University of North Carolina Chapel Hill, North Carolina, United States
Introduction: Digitally reconstructed radiographs (DRRs) are x-ray projections generated from 3D CT scans. These can be leveraged to solve a number of problems that spine surgeons face in the preoperative and intraoperative phases. This work serves to demonstrate the utility of DRRs when combined with machine learning techniques for the specific application of intra-operative level localization. This work aims to provide the surgeon with automatically computed exact level localization, in real-time, using a single fluoroscopic shot.
Methods: Utilizing the VerSe: Large Scale Vertebrae Segmentation Challenge dataset of 374 labeled spine CTs, we generated AP and lateral DRRs to simulate 100 random c-arm orientations relative to the CT. We simulated the 5 degrees of freedom c-arm projections provide (distance to isocenter, x and y linear offsets, RAO/LAO angle and CRA/CAU angle). Using a single desktop GPU (Nvidia Titan RTX) to accelerate the optimization techniques, we derived the 5 parameters needed by sampling the preoperative CT to recreate each fluoroscopic shot and then measured the error.
Results: 7480 unique fluoroscopic images were simulated from 374 CT scans. Our technique demonstrated an ability to converge to the exact solution 65.7% of the time with an average of 3.9 seconds of computation. The images that converged to the solution were 99.9% accurate despite utilizing single precision computations.
Conclusion : With machine learning techniques, we may be able to provide instant registration and alignment of intra-operative imaging to pre-operative scans. This technique has the potential to improve both the safety and efficiency profile for spine surgery. Further work is needed to account for deformable registration and hyperparameter tuning to account for anatomic shifts that occur as patients are in prone positions.
How to Improve Patient Care: This technique demonstrates how machine learning can leverage preoperative imaging with intra-operative data to provide instant, accurate localization for any portion of the spine. This work may potentially limit the amount of radiation administered to patients and surgical team while improving operative efficiency during the planning and navigational phases of surgery.
